yading@10: /* yading@10: * MLP DSP functions x86-optimized yading@10: * Copyright (c) 2009 Ramiro Polla yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavcodec/mlpdsp.h" yading@10: #include "libavcodec/mlp.h" yading@10: yading@10: #if HAVE_7REGS && HAVE_INLINE_ASM yading@10: yading@10: extern char ff_mlp_firorder_8; yading@10: extern char ff_mlp_firorder_7; yading@10: extern char ff_mlp_firorder_6; yading@10: extern char ff_mlp_firorder_5; yading@10: extern char ff_mlp_firorder_4; yading@10: extern char ff_mlp_firorder_3; yading@10: extern char ff_mlp_firorder_2; yading@10: extern char ff_mlp_firorder_1; yading@10: extern char ff_mlp_firorder_0; yading@10: yading@10: extern char ff_mlp_iirorder_4; yading@10: extern char ff_mlp_iirorder_3; yading@10: extern char ff_mlp_iirorder_2; yading@10: extern char ff_mlp_iirorder_1; yading@10: extern char ff_mlp_iirorder_0; yading@10: yading@10: static const void *firtable[9] = { &ff_mlp_firorder_0, &ff_mlp_firorder_1, yading@10: &ff_mlp_firorder_2, &ff_mlp_firorder_3, yading@10: &ff_mlp_firorder_4, &ff_mlp_firorder_5, yading@10: &ff_mlp_firorder_6, &ff_mlp_firorder_7, yading@10: &ff_mlp_firorder_8 }; yading@10: static const void *iirtable[5] = { &ff_mlp_iirorder_0, &ff_mlp_iirorder_1, yading@10: &ff_mlp_iirorder_2, &ff_mlp_iirorder_3, yading@10: &ff_mlp_iirorder_4 }; yading@10: yading@10: #if ARCH_X86_64 yading@10: yading@10: #define MLPMUL(label, offset, offs, offc) \ yading@10: LABEL_MANGLE(label)": \n\t" \ yading@10: "movslq "offset"+"offs"(%0), %%rax\n\t" \ yading@10: "movslq "offset"+"offc"(%1), %%rdx\n\t" \ yading@10: "imul %%rdx, %%rax\n\t" \ yading@10: "add %%rax, %%rsi\n\t" yading@10: yading@10: #define FIRMULREG(label, offset, firc)\ yading@10: LABEL_MANGLE(label)": \n\t" \ yading@10: "movslq "#offset"(%0), %%rax\n\t" \ yading@10: "imul %"#firc", %%rax\n\t" \ yading@10: "add %%rax, %%rsi\n\t" yading@10: yading@10: #define CLEAR_ACCUM \ yading@10: "xor %%rsi, %%rsi\n\t" yading@10: yading@10: #define SHIFT_ACCUM \ yading@10: "shr %%cl, %%rsi\n\t" yading@10: yading@10: #define ACCUM "%%rdx" yading@10: #define RESULT "%%rsi" yading@10: #define RESULT32 "%%esi" yading@10: yading@10: #else /* if ARCH_X86_32 */ yading@10: yading@10: #define MLPMUL(label, offset, offs, offc) \ yading@10: LABEL_MANGLE(label)": \n\t" \ yading@10: "mov "offset"+"offs"(%0), %%eax\n\t" \ yading@10: "imull "offset"+"offc"(%1) \n\t" \ yading@10: "add %%eax , %%esi\n\t" \ yading@10: "adc %%edx , %%ecx\n\t" yading@10: yading@10: #define FIRMULREG(label, offset, firc) \ yading@10: MLPMUL(label, #offset, "0", "0") yading@10: yading@10: #define CLEAR_ACCUM \ yading@10: "xor %%esi, %%esi\n\t" \ yading@10: "xor %%ecx, %%ecx\n\t" yading@10: yading@10: #define SHIFT_ACCUM \ yading@10: "mov %%ecx, %%edx\n\t" \ yading@10: "mov %%esi, %%eax\n\t" \ yading@10: "movzbl %7 , %%ecx\n\t" \ yading@10: "shrd %%cl, %%edx, %%eax\n\t" \ yading@10: yading@10: #define ACCUM "%%edx" yading@10: #define RESULT "%%eax" yading@10: #define RESULT32 "%%eax" yading@10: yading@10: #endif /* !ARCH_X86_64 */ yading@10: yading@10: #define BINC AV_STRINGIFY(4* MAX_CHANNELS) yading@10: #define IOFFS AV_STRINGIFY(4*(MAX_FIR_ORDER + MAX_BLOCKSIZE)) yading@10: #define IOFFC AV_STRINGIFY(4* MAX_FIR_ORDER) yading@10: yading@10: #define FIRMUL(label, offset) MLPMUL(label, #offset, "0", "0") yading@10: #define IIRMUL(label, offset) MLPMUL(label, #offset, IOFFS, IOFFC) yading@10: yading@10: static void mlp_filter_channel_x86(int32_t *state, const int32_t *coeff, yading@10: int firorder, int iirorder, yading@10: unsigned int filter_shift, int32_t mask, yading@10: int blocksize, int32_t *sample_buffer) yading@10: { yading@10: const void *firjump = firtable[firorder]; yading@10: const void *iirjump = iirtable[iirorder]; yading@10: yading@10: blocksize = -blocksize; yading@10: yading@10: __asm__ volatile( yading@10: "1: \n\t" yading@10: CLEAR_ACCUM yading@10: "jmp *%5 \n\t" yading@10: FIRMUL (ff_mlp_firorder_8, 0x1c ) yading@10: FIRMUL (ff_mlp_firorder_7, 0x18 ) yading@10: FIRMUL (ff_mlp_firorder_6, 0x14 ) yading@10: FIRMUL (ff_mlp_firorder_5, 0x10 ) yading@10: FIRMUL (ff_mlp_firorder_4, 0x0c ) yading@10: FIRMULREG(ff_mlp_firorder_3, 0x08,10) yading@10: FIRMULREG(ff_mlp_firorder_2, 0x04, 9) yading@10: FIRMULREG(ff_mlp_firorder_1, 0x00, 8) yading@10: LABEL_MANGLE(ff_mlp_firorder_0)":\n\t" yading@10: "jmp *%6 \n\t" yading@10: IIRMUL (ff_mlp_iirorder_4, 0x0c ) yading@10: IIRMUL (ff_mlp_iirorder_3, 0x08 ) yading@10: IIRMUL (ff_mlp_iirorder_2, 0x04 ) yading@10: IIRMUL (ff_mlp_iirorder_1, 0x00 ) yading@10: LABEL_MANGLE(ff_mlp_iirorder_0)":\n\t" yading@10: SHIFT_ACCUM yading@10: "mov "RESULT" ,"ACCUM" \n\t" yading@10: "add (%2) ,"RESULT" \n\t" yading@10: "and %4 ,"RESULT" \n\t" yading@10: "sub $4 , %0 \n\t" yading@10: "mov "RESULT32", (%0) \n\t" yading@10: "mov "RESULT32", (%2) \n\t" yading@10: "add $"BINC" , %2 \n\t" yading@10: "sub "ACCUM" ,"RESULT" \n\t" yading@10: "mov "RESULT32","IOFFS"(%0) \n\t" yading@10: "incl %3 \n\t" yading@10: "js 1b \n\t" yading@10: : /* 0*/"+r"(state), yading@10: /* 1*/"+r"(coeff), yading@10: /* 2*/"+r"(sample_buffer), yading@10: #if ARCH_X86_64 yading@10: /* 3*/"+r"(blocksize) yading@10: : /* 4*/"r"((x86_reg)mask), /* 5*/"r"(firjump), yading@10: /* 6*/"r"(iirjump) , /* 7*/"c"(filter_shift) yading@10: , /* 8*/"r"((int64_t)coeff[0]) yading@10: , /* 9*/"r"((int64_t)coeff[1]) yading@10: , /*10*/"r"((int64_t)coeff[2]) yading@10: : "rax", "rdx", "rsi" yading@10: #else /* ARCH_X86_32 */ yading@10: /* 3*/"+m"(blocksize) yading@10: : /* 4*/"m"( mask), /* 5*/"m"(firjump), yading@10: /* 6*/"m"(iirjump) , /* 7*/"m"(filter_shift) yading@10: : "eax", "edx", "esi", "ecx" yading@10: #endif /* !ARCH_X86_64 */ yading@10: ); yading@10: } yading@10: yading@10: #endif /* HAVE_7REGS && HAVE_INLINE_ASM */ yading@10: yading@10: av_cold void ff_mlpdsp_init_x86(MLPDSPContext *c) yading@10: { yading@10: #if HAVE_7REGS && HAVE_INLINE_ASM yading@10: c->mlp_filter_channel = mlp_filter_channel_x86; yading@10: #endif yading@10: }